Real-Time Syst (2007) 37: 261–289DOI 10.1007/s11241-007-9025-2

Robust and timely communication over highly dynamicsensor networks

Tian He · Brian M. Blum · Qing Cao ·John A. Stankovic · Sang H. Son ·Tarek F. Abdelzaher

Published online: 29 August 2007© Springer Science+Business Media, LLC 2007

Abstract Highly dynamic sensor networks, such as mobile robotic sensor networks,have been applied in various kinds of application scenarios such as real-time planetexploration and deep-ocean discovery. In these types of networks, mobility and en-ergy management protocols change the connectivity among the neighboring nodesquickly. Traditional state-based protocols, designed for static and/or low-mobilitynetworks, suffer excessive delay in updating their routing or neighborhood tables,leading to severe packet loss and communication delay in the highly dynamic sit-uations. To provide robust and timely communication, we exploit the concept ofLazy-Binding to deal with the elevated network dynamics. Based on this concept andthe knowledge of the node positions, we introduce Implicit Geographic Forwarding(IGF), a new protocol for highly dynamic sensor networks that is altogether state-free. We compare our work against several typical routing protocols in static, mobileand energy-conserving networks under a wide range of system and workload con-figurations. In the presence of mobility and other dynamics, IGF achieves as muchas 10 times improvement in the delivery ratio and significant reduction in both theend-to-end delay and control overhead. In addition to extensive simulations, we alsoimplement and evaluate the IGF protocol on the Berkeley mote platform.

Keywords Wireless sensor networks · Mobile network · Energy conservation

T. He (�)Department of Computer Science and Engineering, University of Minnesota, Minneapolis,MN 55455, USAe-mail: tianhe@cs.umn.edu

B.M. Blum · J.A. Stankovic · S.H. SonDepartment of Computer Science, University of Virginia, Charlottesville, USA

Q. Cao · T.F. AbdelzaherDepartment of Computer Science, University of Illinois, Urbana-Champaign, USA

262 Real-Time Syst (2007) 37: 261–289

1 Introduction

Highly dynamic sensor networks, such as mobile robotic sensor networks, have beenwidely used to explore environments that are difficult and dangerous for humans.In the exploration missions of the red planet, scientists employ the robotic sensordevices (Rover) to discover the possibility of water activity. In the deep-ocean ex-ploration, robotic sensors are used to access the risk of potential earthquakes, volca-noes and tsunamis. In addition, robotic sensors are used to investigate dangerous sitessuch as radioactive environments and mine fields. These robotic sensors are normallyequipped with various kinds of sensors such as cameras, spectrometers and magne-tometers. They can localize themselves in real-time through either GPS (Wellenhoffet al. 1997) or position tracking (Borenstein et al. 1996). Distributed data process-ing, such as the collaborative exploration for map construction (Burgard et al. 2000),requires a team of robotic sensors to communicate with each other constantly inreal-time. This type of communication imposes several challenges. First, when ro-botic sensors move, the constant changes of the connectivity among the neighbor-ing nodes make it difficult to maintain freshness of the routing states in traditionalstate-based routing protocols. Second, energy management protocols (He et al. 2004;Yan et al. 2003; Ye et al. 2003) transit the robotic sensors into and out of sleep states,and their participation in the network becomes probabilistic at any given point intime. These unique challenges demand a new routing solution. In this paper, weexploit the concept of lazy-binding, which is widely used in other research areas,such as the programming language design and operating systems. Specifically here,we define lazy-binding as deferring mapping the system physics (e.g., the networktopologies) into the volatile states (e.g., the route state), required by a certain op-eration, to the last possible moment allowed by the operations. Since lazy bindingdefers binding the volatile states as late as possible, it enables the system to copewith real-time changes in the network topology. Our first installment based on thisconcept is a location-based routing protocol, called Implicit Geographic Forwarding(IGF). IGF allows a sender to determine a packet’s next-hop online in real-time. Bycombining lazy-binding and location-address semantics, IGF becomes a pure state-free protocol, which does not depend on the knowledge of the network topology orthe presence/absence of other nodes. This characteristic of being state-free is valu-able to the highly dynamic sensor networks, as it supports fault tolerance and makesprotocols robust to real-time topology shifts or node state transitions. Further, a state-free solution eliminates the bandwidth-consuming packets required in the state-basedsolutions for routing and neighbor table upkeep.

The remainder of the paper is organized as follows: Sect. 2 motivates the needfor lazy-binding in highly dynamic networks. Section 3 describes the IGF protocolin detail. Section 4 presents our simulation experiments and analysis in mobile andother environments. Section 5 describes our implementation on the MICA2 platformand its evaluation. We discuss the state-of-the-art and future work in Sects. 6 and 7,respectively. We conclude the paper in Sect. 8.

Real-Time Syst (2007) 37: 261–289 263

2 The motivation for lazy-binding

Advances in the protocol design (Intanagonwiwat et al. 2000; Broch et al. 2001;Perkins and Royer 1999; Perkins and Bhagwat 1994; Karp and Kung 2000) contin-uously expand our ability to deal with the high dynamics inside the network. Theseprotocols have been designed with robustness in mind, however, the level of fault tol-erance is usually designed to adapt to occasional node failures and infrequent topol-ogy migration. In order to cope with the elevated transition of network topologies, thestate-based solutions are required to refresh the routing states in real time to reflectchanges, which consequently introduces significant overhead and network conges-tion. Eventually, the performance of these algorithms might degrade dramatically, asthe real-time maintenance of the routing states might not keep up with the transitionrate of the network topologies. We observe that the delay between the time whena physical network topology maps to the routing states and the time when thesestates are actually used for packet forwarding is the root cause of state invalidationand routing failures. We term this delay as the binding delay. A long binding delayleads to a high probability that recorded states are invalid by the time they are used.This problem increases as the network dynamics increase. In addition, since rout-ing states are volatile and become outdated at a much faster rate in highly dynamicnetworks, it is inefficient to maintain state proactively and eagerly. According to thebinding time, we categorize the routing protocols into four categories as shown inFig. 1.

• Fixed routing schemes are rarely used due to their rigid early-binding at the de-ployment time. The binding delay of this type of network could be infinite.

• Proactive schemes such as DSDV (Perkins and Bhagwat 1994) and GPSR (Karpand Kung 2000) maintain the network states aggressively. The routing states arerefreshed regardless whether there is need of data delivery. This eager and proactivebinding property is suitable for the networks with a small rate of topology change.The binding delay is the interval between consecutive routing updates.

• On-demand algorithms such as DSR (Broch et al. 2001), AODV (Perkins andRoyer 1999) and Directed Diffusion (Intanagonwiwat et al. 2000) bind the routingstate to physical topologies with a lazier approach—On-demand Route Discovery.The on-demand property allows them to defer the binding of the routing states tothe physical network topologies until there is a need for end-to-end delivery. Thoseschemes have been proven effective (Broch et al. 1998) in dealing with moderatemobility and failures. The binding delay of on-demand algorithm is the time sincethe route discovery.

Fig. 1 Difference in binding time

264 Real-Time Syst (2007) 37: 261–289

• Different from the on-demand schemes, the IGF protocol proposed in this workgoes one-step further. It defers the binding of the routing states to the physicalnetwork topology until individual packet forwarding operation actually happensat a sending node. This key design difference allows: (1) The elimination of thecommunication overhead to maintain the state proactively, reducing the unneces-sary update of the volatile routing states, (2) The real-time detection of the nodefailure, migration, and transition into a sleep state and (3) The real-time utilizationof recently awoken or newly arriving nodes.

3 IGF protocol design

In this section, we introduce the IGF protocol as an exemplar instance of the lazy-binding concept applied to routing.

3.1 System model and assumptions

IGF is targeting to the high-end sensor networks (e.g., mobile sensor networks),where each sensor node can obtain its location (x, y) through GPS (Wellenhoff et al.1997) or a position tracking technique (Borenstein et al. 1996). The IGF commu-nication supports the location-address semantic, in which locations are specified asthe routing destinations, instead of using a particular node ID. This location-addresssemantic are valid in many sensor networks, because sensor data, such as temper-ature readings, are normally tagged with the location-context, and therefore can beaddressed directly by the location, eliminating the overhead to translate the targetdestinations into a set of node IDs. Since the packet size in high-end sensor networksis relatively large, our main design uses RTS-CTS handshaking to avoid the hiddenand exposed terminal problems in wireless communication (Bharghavan et al. 1994;IEEE 802.11 Group 1997), and an alternative solution for small-packet delivery (e.g.,Tinyos Message) is discussed separately in Sect. 3.8.4. For the sake of simplicity, wedescribe IGF in Sect. 3.2 assuming a sufficient node density. The issues related to thedensity, radio irregularity and localization error are resolved in the later sections.

3.2 IGF details

We begin our introduction of IGF with an straightforward example. Figure 2 depictsa scenario where the node S is transmitting a packet towards the final destinationD. We define the dark nodes within the 60-degree sector1 as forwarding candidates(We address the case when there are no candidates inside the specified forwardingarea in Sect. 3.7). Among these candidates, we highlight two nodes, R and A, torepresent the chosen next-hop and an alternate “competing” node, respectively. Inaddition, the gray node N represents a node within the communication range of S

that is not a candidate node. When the node S initiates a packet transmission, thecommunication handshake goes through following steps: (the timeline of the IGFHandshake is shown in Fig. 3).

1We choose this 60-degree sector because it is the maximal angle that ensures the distance between anytwo points within the sector is smaller than nominal communication range (for the purpose of overhearing).In case of high node density, a smaller angle can be chosen as well.

Real-Time Syst (2007) 37: 261–289 265

Fig. 2 Forwarding area for source S

Fig. 3 IGF handshake

(1) ORTS PHASE: With modifications to the 802.11 DCF MAC protocol, the IGFhandshake begins when the sender S’s Network Allocation Vector (NAV) timeris zero and it senses an idle channel for DIFS time; at this point the node S

sends, via broadcast, an Open RTS (ORTS) packet. This ORTS packet containsthe locations of the sending node S and the final destination D.

(2) CTS-WAIT: While all nodes within the communication radius of the node S re-ceive and process this ORTS packet, only the forwarding candidates (dark nodes)set a CTS_Response timer (Tcts_wait). This timer controls an appropriate amountof time that a forwarding candidate must wait before responding to the receivedORTS packet. The value of Tcts_wait can depend on the link quality, the progressin distance towards the destination and/or the energy remaining at the potentialreceiver. The nodes that are not forwarding candidates (gray nodes) set their NAVtimer in accordance with 802.11 semantics to avoid interference with this ongo-ing transmission.

(3) CTS: While all forwarding candidates set their CTS_Response timer, only a sin-gle node, with the shortest Tcts_wait value (the node R in the Fig. 2 scenario),responds to the ORTS with a CTS packet. To prevent multiple responses, other

266 Real-Time Syst (2007) 37: 261–289

forwarding candidates overhearing this CTS packet cancel their timers and settheir appropriate NAV timers. In addition, the sender S, having already receiveda valid CTS packet, ignores further CTS packets heard in response to the nowantiquated ORTS. We consider the IGF lazy-binding done, when the sender S

decides that the node R is the receiver for this packet.(4) DATA: After the sender S is bound with a specific receiver (R), the sender S

sends DATA to the node R.(5) ACK: The node R acknowledges the sender S, if DATA is received successfully.

3.3 More on forwarding candidates

This section gives a detailed discussion on how a node determines whether it is a validforwarding candidate. We prefer that every node within the forwarding area is capableof hearing one another, to prevent the interference among the forwarding candidates.Accordingly, as depicted in Fig. 2, we choose the candidate nodes that reside withina ±30-degree angle of the line connecting the sender and the final destination.

Using the sender and the receiver’s own location, as well as the final destinationlocation, each node (e.g., the node R) apply simple trigonometry to test whether itselfis within the forwarding area. The formula to calculate the angle ∠RSD in Fig. 2 is:

Degree∠RSD = acos

( |SR|2 + |SD|2 − |RD|22|SR||SD|

). (1)

In the ideal case, the shape of the forwarding area ensures all forwarding candi-dates, responding to an ORTS packet, are located within the communication range ofone another; this eliminates the chance multiple CTS are sent in response to a singleORTS packet. However, in reality, due to an irregular communication radius (Zhouet al. 2004), the node A might still fail to know that a response to the node S’s ORTShas already been transmitted by the node R. In this rare case, the sender S needs toresolve duplicate CTS packets by choosing only one of those responses.

As stated before, the neighboring nodes that receive an ORTS, but are not withinthe forwarding area, simply set their NAV timer to reflect the duration of commu-nication. This prevents collisions due to the hidden terminal problem (IEEE 802.11Group 1997).

3.4 More on setting response wait times

This section provides more discussion on how to set the Tcts_wait value. Having de-termined that it is within the forwarding area of communication, a node can adoptdifferent policies in setting its Tcts_wait according to any combination of availablemetrics including the reception quality of the link, the progress in distance towardthe destination, the energy remaining at the receiver, the statistics of packet loss, theprocessor load or the single hop delay. While many metrics can be used to decide theTcts_wait delay according to the application specifics, without loss of generality, in the

Real-Time Syst (2007) 37: 261–289 267

current IGF implementation, we adopt following formula:

F = WP ∗ (1 − progress/radius) + WR ∗ rand( )

WP + WR

,

Tcts_wait = SIFS + (DIFS − SIFS) ∗ rand( ).

(2)

In (2), progress is the advance in distance toward the destination; Radius is thenominal radio range. Rand( ) generates a random number between 0 and 1; WP andWR are the weights of progress and randomness, respectively; A higher WP tends tofavor a few nodes that make most progress, leading to short routes, however, possiblyless traffic distribution and lower link quality. A higher WR distributes the traffic tomore nodes, leading to a better load balance, however, possibly longer routes. Thetradeoff between WP and WR depends on the link quality and traffic distribution. Weleave the investigation on this tradeoff as future work. SIFS delay is the Short InterFrame Spacing and DIFS delay is the Distributed Inter Frame Spacing as defined inthe 802.11 standard. Equation (2) probabilistically allows the nodes that relay packetsfurther to wait for a smaller period of time before responding. In addition, the ran-domization included in (2) can disperse the system workload among multiple equallyeligible nodes. It should be noted that (2) is designed to be compatible with the tim-ing rule of 802.11 DCF by guaranteeing: (1) The minimum value of Tcts_wait is largerthan or equal to the SIFS delay. (2) The maximum value of Tcts_wait is smaller thanthe DIFS delay to prevent other nodes from initiating a new transmission.

3.5 More on identifying a unique candidate

If more than two forwarding candidates choose similar Tcts_wait values (within prop-agation delay τ ), the transmission of CTS would overlap each other, leading to col-lision. This section provides analysis on the chance of collision under different nodedensities. Here we use the time slotted approach (e.g. in ALHOA and CSMA) toanalyze the performance of the contention-based protocols and establish a systemmodel. The analytic result from the slotted approach serves as the worst-case boundof the un-slotted case. Let Nnode be the average number of competing nodes withinthe forwarding area and Kslot be the number of back-off slots. A CTS packet en-counters a collision when it overlaps with the transmission of at least one other CTSpacket from other competing nodes (two or more CTS packets choose the same slot).A unique candidate can be identified as long as the sender receives at least one CTSresponse from any node within the forwarding area. According to the generalizedbirthday problem (Wagner 2002), the expected number of slots containing exactlyone CTS packets with Nnode competing CTS packets is:

E(Nnode) = Nnode

(1 − 1

Kslot

)Nnode−1

. (3)

According to (3), we plot E(Nnode) values under different Nnode and Kslot settings.Figure 4 suggests that the collision-free slots increase almost linearly with the totalslots available.

We also simulate the process to identify a unique candidate. Figure 5 shows theprobability of success under different Nnodes and Kslots values. Figure 5 indicates that

268 Real-Time Syst (2007) 37: 261–289

Fig. 4 Expected number of slots that are collision-free

Fig. 5 Success ratio in identifying unique candidates

Real-Time Syst (2007) 37: 261–289 269

with a sufficient and reasonable number of back-off slots (e.g., 20) the success ratioapproaches 100%.

Although the chance of collision is very small under our design, it is still possiblewe can not identify a candidate uniquely in some cases. In this case, the standard802.11 time-out mechanism will be activated to retransmit a RTS message.

3.6 About lazy binding in IGF

IGF is an extension of the location-based protocols with the addition of lazy binding.In the location-based protocols such as GPSR, routing depends on up-to-date localneighborhood tables. Normally the neighbor table is updated through periodic bea-coning. The binding of a specific forwarding node to a certain geographic locationis eagerly established when the neighboring nodes exchange beacons. This eager-binding would be invalid quickly due to node mobility or sleep state transitions,which lead to stale routing information and unnecessary beacon exchanges. More-over, this eager binding is not synchronized with the packet forwarding operations.In GPSR, AODV, DSR, if the chosen forwarding node of a sender fails or moves outof range, the MAC layer of the sender drops the packet and notifies the network layerabout the routing failure. The network layer has to resolve this failure by attemptinganother backup route if available, which might be invalid too, or alternately waitingfor an update to the neighborhood table, suffering a latency proportional to the bea-con period in a scale of seconds. In contrast, IGF adopts lazy binding to discover thenext hop the instant it is needed. The worst-case back-off delay introduced by lazybinding is tens of microseconds according to the 802.11a standard. This is four ordersof magnitude shorter than the period of the neighbor table update through beaconingfound in other protocols. The worst-case back-off delay in our implementation onthe MICA platform is higher due to a low data rate; however, it is still two ordersof magnitude shorter. In addition, the route maintenance found in Directed Diffusion(Intanagonwiwat et al. 2000), DSR (Broch et al. 2001), AODV (Perkins and Royer1999) and LAR (Ko and Vaidya 1998) normally takes at least tens of millisecondsor seconds to fix a broken link (depending on the size of network and the cause offailure). In contrast, IGF binds the node that is able to forward the packets, momentsbefore (about 50us) the actual forwarding operation takes place. This lazy bindingproperty dramatically reduces the chance that packets are forwarded to a node thatfails or moves out of range. As a result, IGF shows as much as 10 times performanceimprovement in the delivery ratio when compared with several classical and state ofthe art solutions in the presence of high rate changes of the network topology.

3.7 Optimizations for sparse networks

IGF targets sensing-covered dense sensor networks in which greedy forwarding hasbeen proven to guarantee delivery (Xing et al. 2004). However, we note that with-out the capability of circumventing voids (e.g., the absence of forwarding candidatenodes), IGF results in communication failure in sparse sensor networks. The state-less property of IGF precludes utilizing the perimeter-forwarding rule for planarizedgraphs, such as the one used in GPSR (Karp and Kung 2000).

270 Real-Time Syst (2007) 37: 261–289

Fig. 6 IGF handshake

To improve the delivery performance under sparse sensor networks, we have de-signed and implemented a history-based forwarding area shift technique in IGF. Thismechanism activates when a void is detected through a MAC layer notification offailure to IGF. The sender then retransmits the packet, requesting a shift of the for-warding area to search for an available receiver. The sequence of shifts is shownFig. 6. Those shifts allow IGF to utilize communication areas outside of the initialforwarding area. Since the area shifts allow backtracking, we must make sure thatIGF is loop-free while maintaining the state-free property. Unfortunately, it has beenproven in Stojmenovic and Lin (1999), a memoryless location-based routing algo-rithm is not loop-free if backtracking is allowed. To address this issue, IGF placesa trace-history into the packet header to remember the nodes this packet visited re-cently, and no state is maintained in the nodes. To avoid the infinite loop, during thebacktracking, a node choose the next hop forwarding node with an ID that is not inthe trace-history. We note that this trace-history starts to accumulate only when thebacktracking is activated, it does not incur overhead whenever greedy forwarding ispossible. With this void avoidance capability, in the empirical study later shown inthe evaluation, IGF is able to achieve a 100% delivery ratio with a small length ofhistory added to a packet header.

3.8 Design issues

This section completes our approach with several practical design issues.

3.8.1 Radio irregularity

For the sake of clarity, IGF is described with a nominal symmetric radio range. How-ever, IGF does work with asymmetric irregular range (Zhou et al. 2004). First, weenforce a symmetric channel by an ORTS-CTS-DATA-ACK handshaking procedure;second, though it is possible that an asymmetric channel among forwarding candi-dates still exists, which might introduce multiple CTS responses to a single ORTS,the sender can resolve the duplicate packets by simply choosing one and ignoring theothers.

Real-Time Syst (2007) 37: 261–289 271

3.8.2 Localization error impact

IGF can be regarded as an extension to location-based protocols, whose performancecan be affected by localization errors. Results from He et al. (2003a) show that theperformance of the Geographic Forwarding (GF) protocol degrades as the localiza-tion error increases. In our evaluation Sect. 4.5, we demonstrate our IGF scheme aswell as GPSR achieves 100% delivery ratios in the presence of up to 50% radio rangeerrors.

3.8.3 Energy implications

The 802.11b standard allows a node to turn off the radio (Ye et al. 2002b) after thenode overhears a RTS packet that is not targeted to itself. However, IGF requiresforwarding candidates to remain in the listening mode to overhear any CTS for about5 × 10−5 seconds. This causes a slightly increase in the energy consumption. Wenote, however, that this increase is negligible when considering the higher deliveryratio, the reduction in control packets, and the smaller end-to-end delay we are ableto achieve.

3.8.4 Alternative MAC implementation

Without loss of generality, IGF is currently built and evaluated on the top of 802.11DCF. Considering the bandwidth available in the mobile robotic sensor networks, it isa good solution in dealing the hidden and exposure terminal problems. However, wenote that IGF is not bound to 802.11 DCF and it can be implemented with several ex-isting MAC protocols. For example, IGF can be built on the MAC protocol suggestedin Woo and Culler (2001) that uses the implicit ACK. In this scenario, forwardingcandidates wait for a random delay before starting to relay a packet, and the one withthe smallest delay forwards the data packet first. This data packet serves as both anacknowledgement to the sender and a cancellation signal to the rest of the forwardingcandidates. (These handshaking sequences are shown in Fig. 7.) Based on Wan et al.(2002), we have built the IGF protocol on the Berkeley mote platform and evaluatedit with results shown in Sect. 5.

Fig. 7 IGF handshake

272 Real-Time Syst (2007) 37: 261–289

3.8.5 Other implications

Depending on the localization method used, IGF requires either additional energy(Wellenhoff et al. 1997) or more control messages (He et al. 2003a) to localize thenodes, especially in the mobile environments. In addition, IGF doesn’t fix the routesduring the forwarding, which might lead to packet reordering due to the MAC con-tention. Consequently, the final receiver should ensure the data can be re-assembledcorrectly.

4 Experiments and analysis

To assess the performance, we implement the IGF protocol in GloMoSim (Zeng et al.1998), a simulator for wireless sensor, ad hoc, and mobile networks. GloMoSim pro-vides a high fidelity simulation for wireless communication with detailed modeling ofcommunication propagation, radio and MAC layers. In addition, we also implementthe IGF protocol on Berkeley mote platform (Sect. 5).

To make our evaluation close to the latest Telos mote capability proposed for usein the WSN environments (CrossBow, Telos Mote IEEE 802.15.4 Compliant MoteData Sheet, http://www.moteiv.com/products/telos-brochure.pdf), we set our systemparameters as shown in Table 1. We expect the typical communication patterns insidea sensor network to be established based on request and retrieval semantics for datadelivery between sensor nodes and a querying entity. One-to-one, many-to-one andmany-to-many communication patterns are representative workloads in sensor net-works. One-to-one communication happens when one node detects some activity thatneeds to be reported to a remote entity. Alternatively, a querying entity will requireperiodic reports from the whole sensor area, which take the form of many-to-onecommunication. It is more common that multiple applications run simultaneouslyand the traffic flows interleave with each other, shown by the many-to-many cross-traffic pattern.

We evaluate 120 system configurations under different traffic loads, node mo-bility and energy conserving schedules. For each configuration we average 60 runswith different random seeds (hence 60 different network topologies and node place-ments) to ensure adequate confidence of our results. The 90% confidence intervalis within 3% to 10% of the mean for GPSR and IGF, and 8% to 15% for LAR.

Table 1 System parametersParameters Settings

Radio Range 40.0 m

Terrain 150 × 150 m2

Collision Range 71.2 m

Nodes 100 nodes, Uniform

Bandwidth 200 kbps

Radio Lossy channel

Packet Size 32 byte Payload

WP WP = 2

WR WR = 1

Real-Time Syst (2007) 37: 261–289 273

Due to the space limitation, we only present more complex and interesting many-to-many scenario (40 × 60 = 2400 runs). The complete data set is available uponrequest. In the many-to-many tests, 6 nodes, randomly chosen from the left sideof the terrain, send 6 CBR flows to 2 nodes (3 each) on the right side of the ter-rain. The average hop count is about 4−6 hops. We note that most well-knownsensor network protocols such as Directed Diffusion (Intanagonwiwat et al. 2000),TTDD (Ye et al. 2002a) and TBF (Niculescu and Nath 2003) are mostly designedfor static sensor networks and have never been evaluated in mobile environments.For the sake of fairness, we choose the only protocols that evaluate the mobil-ity extensively in their publications (Broch et al. 1998, 2001; Karp and Kung 2000;Ko and Vaidya 1998). Moreover, since IGF is a location-based routing protocol, it isunfair to compare IGF with other ID-based protocol. As a result, we decide to com-pare IGF with two protocols: (1) LAR (Ko and Vaidya 1998) is a protocol optimizedfor mobility using the location information and it is suitable for sensor network; and(2) GPSR (Karp and Kung 2000) is the standard location-based sensor network pro-tocol with greedy and planar perimeter forwarding rules. We consider these protocolsin three scenarios:

• A Static Network, where nodes are not mobile and energy conservation is not con-sidered;

• A Mobile Network, with mobility ranging from walking to vehicular speeds;• An Energy Conservation Network where nodes can transition into and out of dor-

mant states.

For each experiment we choose three typical metrics on (1) the delivery ratio (thenumber of packets received/number of packets sent), (2) average end-to-end delayof received packets,2 and (3) overall communication overhead (total packets sent outby a node). In addition to the above experiments, we also evaluate the performancesensitivity of IGF in the presence of a low node density (voids), localization errorsand location update delay in Sects. 3.7, 4.5 and 4.6, respectively.

4.1 Performance in static networks

The evaluation in static networks shows that IGF performs as well as or slightly betterthan GPSR and LAR, when dynamics such as mobility and energy conserving sleepcycles are not considered. In these experiments, we increase many-to-many CBRflow rate until sufficient congestion is seen. Figure 8a shows that GPSR and IGFhave comparable delivery ratios under light loads, while LAR loses packets earlyas these protocols quickly congest the network by sending route discovery packets.When the traffic flow rates increase enough to adequately congest the network inGPSR and IGF (6+ packets/second per CBR flow), performance in GPSR degradesdue to limited intersecting routes, suffering additional collision caused by neighbortable update beacons (0.1 beacon per second). LAR uses location information to keepthe effects of route discovery to a minimum allowing it to maintain delivery ratios

2We note that our evaluation does not choose deadline miss ratios as the major metrics, because such anapproach reveals less information about the tradeoff between actual delays and other system performanceparameters.

274 Real-Time Syst (2007) 37: 261–289

Fig

.8Pe

rfor

man

cein

stat

icne

twor

ks

Real-Time Syst (2007) 37: 261–289 275

comparable to IGF. However, LAR’s frequent transmission of route discovery packetstoward the destination, coupled with the latency incurred awaiting the route discoveryresponse, lead to significantly more overhead (Fig. 8b) and longer delay (Fig. 8c)when compared with IGF. Figure 8b demonstrates IGF’s savings at low traffic loads,as IGF does not require beaconing (GPSR) or route discovery packets (LAR) requiredin these protocols. As traffic loads increase, congestion increases the number of MAClayer collisions in both IGF and GPSR, resulting in retransmission attempts that addto overhead as shown in Fig. 8b. In GPSR and IGF we see significantly lower end-to-end delay beyond 4 packets/second per CBR flow because LAR suffers latencyawaiting the return of route discovery packets. Finally, under heavy traffic, we seea slightly longer delay in IGF over GPSR due to the fact that IGF manages to deliver10% more packets (Fig. 8a). Analysis in Sect. 3.5 indicates that more back-off slotsare needed for highly dense networks. However, since the duration of a back-off slotis much smaller than the delay caused by queuing and retransmission, the CTS back-off delay due to lazy binding in IGF has virtually no impact on the end-to-end delayeven at a density of 22 node/per nominal radio range as shown in Fig. 8c.

4.2 Performance under mobility

One scenario for the evaluation under high mobility would be a group of exploringrobotic sensor nodes, trying to find the survivors underneath rubble after an earth-quake. They periodically update their locations to each other while searching. An-other scenario would be a group of mobile robots equipped with magnetic sensors,searching for mines in a battlefield. These robots report the detections to a base sta-tion by relaying packets among themselves. As we mentioned before, nodes’ loca-tions can be obtained through GPS in such a highly dynamic systems. We choosea standard waypoint mobility model during the simulations. It should be noted thatin contrast to ad hoc networks where the mobility pattern is interleaved with burstmovements and long pauses, sensor robots are normally continuously moving. Toreflect this scenario, we set only a 1-second pause interval between moves (100–1000 s pause intervals are normal settings in ad hoc network evaluations Wellen-hoff et al. 1997). The settings stress-test the protocols’ ability to deal with contin-uously high mobility and reflect the mobility patterns seen in mobile sensor net-works. We model speeds up to 18 meters per second (∼40 mph) to evaluate a widerange of mobile scenarios in which sensors can be attached to slow robots or tohigh-speed vehicles. We adopt a 40 m range to confirm to current sensor ability,which is much smaller than 250 m setting in used in WLAN (Wellenhoff et al. 1997;Zhou et al. 2004). We note that the mobility is characterized by the number of neigh-borhood changes per second, which is affected by both the node speed and the radiorange. With the same speed, a smaller range leads to a high mobility. To validate thispoint, in addition to this section, we investigate the impact of the radio ranges on therouting performance under mobility in Sect. 4.2.1.

In the static network scenario (Sect. 4.1), we use a low beacon rate (0.1 beacon persecond) in GPSR to reduce the effect of congestion. To optimize GPSR to deal withmobility, we test GPSR with multiple beacon rates. Because beacons consume band-width, their cost offsets their savings, arriving at similar results in all beacon rates we

276 Real-Time Syst (2007) 37: 261–289

Fig

.9Pe

rfor

man

cein

mob

ilene

twor

ks

Real-Time Syst (2007) 37: 261–289 277

tested. Consequently, we adopt 1 beacon per second to keep the state as fresh as pos-sible without causing congestion. Rerouting is supported when a protocol experiencea link break. From Fig. 9a, we see that when nodes do not move (0 m/s), no packetsare lost and the lowest delay and overhead are incurred due to minimal congestion. Aswe introduce mobility, increasingly affecting the validity of neighborhood and rout-ing state with increased node speeds, we see the delivery ratios (Fig. 9a) in GPSR andLAR drop off quickly while IGF continues to perform close to optimal. For example,when the node moving speed is at 4 meters/second, IGF demonstrates as much as10 times performance gain in the delivery ratio over GPSR. For LAR, performancedegrades as node migration invalidates eager-binding routes. Since LAR is speciallydesigned to deal with mobility, its milder degradation, as seen in Fig. 9a, results fromlocation controlled flooding of route discovery packets to reestablish routes despitemobility. As an addendum to explain why GPSR performs so poorly, we note fromFig. 9a that GPSR’s delivery ratio quickly drops to zero at relatively low node speeds.One might assume this is because the beacon overhead leads to congestion in GPSR,hence a very low delivery ratio. However, from Fig. 9b, we note that control over-head in GPSR is actually smaller than LAR. In fact, this low delivery ratio in GPSRhappens because according to greedy forwarding rules in GPSR, the chosen next-hopnode is normally located at the edge of the sender’s communication radius. Becausenodes are equally likely to move in any direction, there is a high chance that desig-nated receiver will have moved out of communication range from the sender since thelast beacon which was received seconds ago. Over multi-hop routes, the chances offailure grow exponentially. In contrast, IGF binds the next hop tens of microsecondsbefore packet forwarding occurs. This significantly reduces the chance that a chosennode will move out of communication range during this tiny interval. Aside from thedelivery ratio (Fig. 9a), our evaluation shows that IGF significantly outperforms otherprotocols in metrics of overhead (Fig. 9b) and end-to-end delay (Fig. 9c) under allmoving speeds. All these results are due to IGF’s ability to defer the mapping betweenrouting states and network topologies until this binding is absolutely required.

4.2.1 Radio range impact on the routing performance under mobility

In this experiment, we investigate the impact of different radio ranges on the routingperformance in mobile sensor networks. When nodes move around, mobility breaksold links and establishes new links. With the same node speed, a smaller radio rangecauses a higher rate of change in the network topology. Figure 10 proves that GPSRis able to achieve good deliver ratios with a large radio range, which leads to a smallermobility. On the other hand, Fig. 10 indicates GPSR’s delivery performance reducesdramatically under high mobility situations.

4.3 Performance under energy conservation

It is crucial for sensor network systems to support energy conservation. The mostpractical way to reduce total energy consumption is to turn on/off the nodes on de-mand of events (He et al. 2004; Yan et al. 2003; Ye et al. 2003). However, theseoperations disrupt the network topologies. In this experiment, we test IGF, GPSRand LAR in the presence of orthogonal energy conserving protocols by periodically

278 Real-Time Syst (2007) 37: 261–289

Fig. 10 Delivery ratio under different radio ranges and speeds

transiting nodes into and out of sleep states. To prevent congestion, and therefore iso-late the effects of the awake-sleep transition in our analysis, we set the flow rate to 1packet per second. We note that two key parameters in energy-conserving protocolscan affect the routing performance:

• Toggle period: Toggle Period is the time interval between consecutive transitionsinto a sleep state. This parameter reflects how fast a routing state is invalidated dueto sleep-awake transitions. We change this value from 5 seconds to 95 seconds inincrements of 10.

• Sleep percentages: The percentage of time a node is in the sleep mode. We notethat sleeping can significantly affect the active node density, as this reduces thenumber of nodes participating in routing at any point in time.

4.3.1 Performance under varied toggle periods

Figure 11a shows the results for many-to-many flows where the Sleep Percentageis set at 30% for varying Toggle Periods. It shows that IGF outperforms all otherprotocols at all toggle periods investigated. GPSR utilizes a beaconing mechanism toproactively bind network topologies into neighbor states. This binding can be quicklyinvalidated due to nodes’ awake-sleep transitions. As a result, packets may be for-warded to nodes that were turned off since the last beacon and then dropped by theMAC layer. This leads to a poor delivery ratio in GPSR (Fig. 11a). In LAR, a noderequires the network layer to handle transmission failures by initiating route discov-ery. Due to the on-demand nature of those algorithms, LAR outperform GPSR, as therecently returned route discovery packet traverses nodes that are currently awake andtherefore able to act as routers. Finally, we see IGF performing significantly betterthan other protocols, at times showing more than 3 times improvement in packets de-livered when compared to GPSR. We attribute this performance to the IGF’s abilityto utilize whatever neighbors are currently awake en route to the destination. We notethe Toggle Periods here only range from 5 to 95 seconds. When the Toggle Periods

Real-Time Syst (2007) 37: 261–289 279

Fig

.11

Perf

orm

ance

unde

rva

ried

togg

lepe

riod

s

280 Real-Time Syst (2007) 37: 261–289

increase further, less dynamics are introduced into the network topologies and routingstates can remain fresh for a longer period of time. In this scenario, higher deliveryratios are expected for other algorithms. Theoretically, when the Toggle Period ap-proaches infinity, energy conserving networks become traditional static networks, forwhich we have shown performance comparisons in Sect. 4.1.

4.3.2 Performance under varied sleep percentage

We next assess routing performance varying Sleep Percentage for the highly volatilecase where the Toggle Period is set to 5 seconds. This not only allows us to com-pare our work under varied Sleep Percentage times, but allows us to stress test ourprotocol under highly dynamic system settings. In this experiment, we increase thesleep percentage of each node from 0% (always awake) to 100% (always asleep) inincrements of 10%.

Figures 12a, b and c all demonstrate IGF’s better performance over variedSleep Percentages. Figure 12a shows that IGF delivers the highest percentage ofpackets under all Sleep Percentage settings, while incurring the small end-to-enddelay (Fig. 12c) and the lowest transmission overhead (Fig. 12b). For example,Fig. 12a shows that at a 50% sleep percentage, IGF delivers 340% more packets thanthe GPSR protocol. The drastic drop in overhead (Fig. 12b) as seen in LAR also canalso be attributed to this drop in the Packet Delivery Ratio. Since LAR is designed toadapt to occasional node failures, as we expect, in such highly dynamic networks, ittakes a huge end-to-end delay to repeatedly fix these routes (Fig. 12c). GPSR showsthe lowest end-to-end delay (Fig. 12c) because it delivers a tiny percentage of pack-ets when compared to the IGF. Those packets go through the networks quickly bychance. Only IGF has a highest delivery ratio and a small delay. This is due to thefact that IGF can immediately detect node transitions into sleep states and immedi-ately utilize recently awoken nodes.

4.4 Performance in sparse sensor networks

The typical density of sensing-covered sensor network systems (He et al. 2004) isabout 20–25 nodes/radio range in order to provide high fidelity in localization, detec-tion and tracking. In previous evaluations, we use 22 node/radio ranges as a typicalsetting. However, it is important to understand how IGF performs under various nodedensity settings. To prevent congestion, and therefore isolate the effects of densityin our analysis, we set the per node flow rate to 1 packet per second. To change thedensity of the network, instead of increasing the number of nodes in the terrain, wekeep the number of nodes constant at 100, and increase the side length of the squareterrain from 100 meters to 250 meters in increments of 10. Figure 13 shows that withthe history-based forwarding-area shifts, IGF achieves a 100% delivery ratio whenthe node density is larger than 12 nodes per nominal radio range. Figure 13 revealsthat when density if relatively high (≥9 node/radio range), longer trace-history doesnot help much, however when the network become sparse, longer history can improvethe delivery ratio.

Real-Time Syst (2007) 37: 261–289 281

Fig

.12

Perf

orm

ance

unde

rva

ried

slee

ppe

rcen

tage

282 Real-Time Syst (2007) 37: 261–289

Fig. 13 Delivery ratio vs. nodedensity

Fig. 14 Localization error impact

4.5 Performance under localization errors

While most work in location-based routing assumes perfect location information,the fact is that erroneous location estimates are virtually impossible to avoid. In thisexperiment, we investigate location error impact on the IGF protocol. To prevent con-gestion, and therefore isolate the effects of the localization error, the traffic loads areset to the rate of 1 packet/second. We compare IGF, GPSR with the basic geographicforwarding (GF) (Finn 1987), which forwards a packet to the node that makes themost progress toward the destination. We increase the localization error from 0% to50% of the radio range in increments of 5% to measure the end-to-end delivery ratios.Figure 14 demonstrates that both the IGF and GPSR protocol perform much betterin the presence of localization errors while the GF protocol suffers as location errorsincrease.

4.6 Performance under different localization update intervals

IGF obtains location updates from GPS or other localization schemes. Since the up-date rate affects the amount energy consumed to obtain the locations, the location isnormally updated intermittently. Consequently, nodes have to make the routing de-cisions based on the last localization result, which might cause the routing failures

Real-Time Syst (2007) 37: 261–289 283

Fig. 15 Impact of locationupdate delay

if the update delay is too long. In this section, we investigate the impact of the loca-tion update delay to the end-to-end delivery ratio. Figure 15 shows that the locationupdate delay doesn’t affect the static and energy conservation networks since nodesdon’t move in such networks. As for the mobile networks, a moderate location updatedelay (e.g., ≤1 second) doesn’t noticeably affect the delivery ratio, however, a largedelay cause more routing failures. Figure 15 also indicates that the impact of the up-date interval is affected by nodes’ speed. With the same update intervals, a faster nodespeed leads to a lower delivery ratio.

5 Implementation on motes

We have implemented the IGF protocol on the Berkeley motes platform (Mica2 datasheet 2003) with a code size of 11,606 bytes (code is available through CVS athttp://sourceforge.net/projects/vert/). Currently, this implementation is built on thetop of a MAC protocol with the implicit ACKs mentioned in Sect. 3.8.4. Three ap-plications including data placement, target tracking and CBR data streaming are alsobuilt to run on top of IGF. Due to physical constraints and the un-availability of state-of-the-art protocols on such a platform, it is difficult to perform as extensive evalu-ation as we did in the wireless simulator. We, therefore, only present initial resultshere as a study for developing a more complete solution and evaluation in the futuremote platform. As we mentioned in Sect. 3.2, IGF does not task a specific node toroute packets a priori. This feature is beneficial for load balancing among the nodesinside the forwarding area. In this experiment, we use 25 motes to form a 5 by 5 grid.To evaluate the load balancing capability of IGF we send a CBR data stream fromnode 24 to node 0, which is the base station. We collect the number of packets relayedby intermediate motes (1–23) and compare this with the result obtained from the GFprotocol which we also implemented on the motes. While both GF and IGF achievenearly 100% delivery ratio, GF tends to relay packets via a fixed route which mightlead to unbalanced traffic. This is shown in Fig. 16 as node 19 relays 250 packetswhile node 18 doesn’t forward any packets. Instead, by distributing traffic loads, IGFeffectively balances energy consumption. We argue that in sensor networks, balancedenergy consumption can prevent some nodes from dying faster than others, thereforeincreasing the network lifetime.

284 Real-Time Syst (2007) 37: 261–289

Fig. 16 Traffic balance

6 Related work

In this section we discuss prior research in distributed computing that is pertinentto the design of IGF. Various protocols (Ganesan et al. 2002; Stann and Heidemann2003; Wan et al. 2002, 2003; Woo et al. 2003; Zhao and Govindan 2003) have beenintroduced to reduce packet loss through reliable communication in sensor networks.Alec Woo (Woo et al. 2003) chooses reliable routes based on link connectivity statis-tics obtained dynamically from a EWMA estimator. RMST (Stann and Heidemann2003) tracks packet fragments so that receiver initiated requests can be satisfied whenindividual pieces of an application payload get lost. PSFQ (Wan et al. 2002) cachespackets along the path to the sender, initiating fragment recovery as required, startingwith its local neighborhood. Robust data delivery (Ganesan et al. 2002) simultane-ously sends packets along multiple paths at the expense of increases in communi-cation overhead. While these ARQ/FEC-based solutions have proven effective whendealing with interference and collisions, their robust and reliable features might notbe able to handle failures due to high dynamics in network topologies. We considerthem orthogonal and complementary to our work.

Many routing algorithms have been proposed for ad hoc and sensor networks.With regard to the mechanisms used to bind network topology to the routing state,we divide these routing protocols into three categories. The first category we term asproactive eager-binding routing algorithms. DSDV (Perkins and Bhagwat 1994) re-quires each node to proactively broadcast routing updates periodically. Global routingtables are refreshed regardless whether there is need for data delivery. Location-basedrouting algorithms such as GPSR (Karp and Kung 2000) remove the requirementthat a protocol maintains a global view of the network (i.e. end-to-end routing ta-bles), therefore reduces communication overhead by eliminating its dependence onthe network wide state information. However, they still depend on up-to-date localneighborhood tables, requiring control overhead to maintain such tables and suf-fering latency and packet loss when a node’s neighborhood state changes betweenupdates. To minimize unnecessary overhead incurred by proactive updates, a set of

Real-Time Syst (2007) 37: 261–289 285

on-demand algorithms are proposed to defer route acquisition until data delivery isrequired. We term the second category as reactive eager-binding algorithms. It hasbeen proved in Broch et al. (1998) that AODV (Perkins and Royer 1999) and DSR(Broch et al. 2001) can successfully deal with moderate mobility with long pauseintervals (100–1000 seconds). However, the eager binding of the routing states atthe route acquisition phase make them less effective to deal with high dynamics inwhich network topologies change at a much faster rate than the duration of con-nections. Routing maintenance and rediscovery are proposed in Broch et al. (2001),Perkins and Royer (1999) to remedy this situation partially at the cost of higher delayand expensive control overhead. LAR (Ko and Vaidya 1998) extends the on-demandidea proposed by AODV (Perkins and Royer 1999) and DSR (Broch et al. 2001),utilizing location information to limit the scope of route requests. While LAR signif-icantly reduces routing overhead by only propagating queries to relevant portions ofthe network, it still needs to maintain or establish an explicit path before transmit-ting a packet. Current reactive eager-binding algorithms can successfully deal withoccasional node failures and moderate mobility. However, the elevated dynamics dueto the continuous mobility and power conservation inside sensor networks challengeresearchers to develop a new category of routing protocols based on the lazy bindingconcept.

The first state-free protocol IGF belongs to this third category. ExOR (SanjitBiswas 2005) also decides the forwarding candidate on the fly. However, before trans-mitting a packet, the sender needs to specify the forwarding candidates in the packetheader, which requires maintaining the state information about neighboring nodes.GeRaF (Zorzi and Rao 2003) proposes a similar packet forwarding technique andit focuses on the multi-hop performance in terms of the average number of hops toreach a destination. Both ExOR and GeRaF do not model the effect of channel con-tention; while this work provides a detailed implementation and evaluation throughboth simulation and a running system.

We note that although IGF delivers end-to-end traffic in a timely manner and dealswith real-time topology changes, it is not aiming at providing end-to-end real-timeguarantee. Several real-time protocols have been developed for this purpose. For in-stances, RAP (Lu et al. 2002) uses velocity monotonic scheduling (VMS) to distrib-uted schedule real-time traffic with the assistance from a differentiated MAC Layer.SPEED (He et al. 2003b) uses feedback-based adaptation to enforce per-hop speedin face of unpredictable traffic. As a future work, we will investigate how to augmentIGF with real-time guarantee.

7 Future work

In this work, IGF assumes a localization service or the GPS capability. This is justi-fied as sensor network applications require location information to make sensor datameaningful. We note that lazy-binding is a general concept to deal with high networkdynamics and its applicability does not intrinsically depend on the location service.It is promising to apply lazy-binding to ID-Based protocols such as Directed Diffu-sion (Intanagonwiwat et al. 2000). To extend (Intanagonwiwat et al. 2000), we cankeep the hop-count-to-a-sink as a non-volatile state with respect to the node failures,

286 Real-Time Syst (2007) 37: 261–289

and we perform forwarding operations with the parents of each node. We note that inthis ID-based case, the state-free property is not maintained, however, lazy-binding,which is independent of the state-free property, is still beneficial in dealing with thefailure of the parent nodes. Due to the space constraints, we leave this as future work.

8 Conclusions

In highly dynamic sensor networks, the maintenance of freshness of routing states iscostly. The state update, resulting from eager-binding, directly contributes to networkcongestion, wasting precious energy and increasing the end-to-end transmission la-tency. To prevent the adverse affects that dynamic factors such as high mobility haveon the state-based eager-binding routing protocols, we advocate using the concept oflazy-binding to cope with high dynamics in sensor networks. Based on this concept,we introduce IGF, a unicast protocol that is altogether state-free. In simulation, wecompare our work against protocols designed for mobile environments and sensornetworks. IGF demonstrates more than 10 times improvement in the packet deliveryratio when the sensor network is highly mobile. IGF also achieves significant reduc-tion in delay and overhead when considering mobility and energy-conservation. Inaddition, the IGF protocol has been implemented on the Berkeley motes platform toserve as an initial study in developing a more complete solution in the future.

Acknowledgements This work was supported in part by NSF grant CNS-0626614, CNS-0615063,CNS-0614870, CNS-0626632, CCR-0329609 and CNS-0614886

References

Bharghavan V, Demers A, Shenker S, Zhang L (1994) Macaw: a media access protocol for wireless LAN.In: The SIGCOMM ’94 conference on communications architectures

Borenstein J, Everett H, Feng L (1996) Navigating mobile robots: systems and techniques. AK Peters Ltd,Wellesley

Broch J, Maltz DA, Johnson DB, Hu Y-C, Jetcheva J (1998) Performance comparison of multi-hop wire-less ad hoc network routing protocols. In: The fourth annual ACM/IEEE international conference onmobile computing and networking, 1998

Broch J, Johnson DB, Maltz DA (2001) DSR: the dynamic source routing protocol for multihop wirelessad hoc networks. Ad Hoc Networks, Addison Wesley

Burgard W, Moors M, Fox D, Simmons R, Thrun S (2000) Collaborative multi-robot exploration. In:Proceeding of IEEE international conference on robotics and automation (ICRA), 2000

CrossBow, Telos Mote IEEE 802.15.4 Compliant Mote Data Sheet, http://www.moteiv.com/products/telos-brochure.pdf

Finn GG (1987) Routing and addressing problems in large metropolitan-scale internetworks, USC/ISI,Tech Rep ISI/RR-87-180

Ganesan D, Govindan R, Shenker S, Estrin D (2002) Highly resilient, energy-efficient multipath routingin wireless sensor networks. Mobile Comput Commun Rev (MC2R) 5(4):11–25

He T, Huang C, Blum BM, Stankovic JA, Abdelzaher T (2003a) Range-free localization schemes in large-scale sensor networks. In: MOBICOM’03, September 2003

He T, Stankovic J, Lu C, Abdelzaher T (2003b) SPEED: a stateless protocol for real-time communicationin ad hoc sensor networks. In: ICDCS’03, May 2003

He T, Krishnamurthy S, Stankovic JA, Abdelzaher T (2004) An energy-efficient surveillance system usingwireless sensor networks. In: MobiSys’04, June 2004

IEEE 802.11 Group (1997) Wireless LAN media access control (MAC) and physical control specifications,ANSI/IEEE

Real-Time Syst (2007) 37: 261–289 287

Intanagonwiwat C, Govindan R, Estrin D (2000) Directed diffusion: a scalable and robust communicationparadigm for sensor networks. In: The sixth annual international conference on mobile computing andnetworks, 2000

Karp B, Kung HT (2000) Greedy perimeter stateless routing for wireless networks. In: International con-ference on mobile computing and networking

Ko Y, Vaidya NH (1998) Location-aided routing (lar) in mobile ad hoc networks. In: International confer-ence on mobile computing and networking

Lu C, Blum BM, Abdelzaher TF, Stankovic JA, He T (2002) Rap: a real-time communication architecturefor large-scale wireless sensor networks. In: IEEE RTAS, 2002

Mica2 data sheet, CrossBow, 2003, available at http://www.xbow.com

Niculescu D, Nath B (2003) Trajectory based forwarding and its applications. In: MobiCom, 2003

Perkins CE, Bhagwat P (1994) Highly dynamic destination-sequenced distance-vector routing (dsdv) formobile computers. In: SIGCOMM, 1994

Perkins CE, Royer EM (1999) Ad-hoc on demand distance vector routing. In: WMCSA, 1999

Sanjit Biswas RM (2005) ExOR: opportunistic multi-hop routing for wireless networks. In: Proc of theACM SIGCOMM, August 2005

Stann F, Heidemann J (2003) RMST: reliable data transport in sensor networks. In: SNPA ’03, 2003

Stojmenovic I, Lin X (1999) Gedir: loop-free location based routing in wireless networks. In: Int conf onparallel and distributed computing and systems

Wagner D (2002) A generalized birthday problem (extended abstract). CRYPTO 2002

Wan CY, Campbell AT, Krishnamurthy L (2002) PSFQ: a reliable transport protocol for wireless sensornetworks. In: WSNA ’02, 2002

Wan C, Eisenman S, Campbell AT (2003) CODA: congestion detection and avoidance in sensor networks.In: Sensys 03, 2003

Wellenhoff BH, Lichtenegger H, Collins J (1997) Global positions system: theory and practice, 4th edn.Springer, Berlin

Woo A, Culler D (2001) A transmission control scheme for media access in sensor networks. In: Proc. ofmobile computing and networking (Mobicom), 2001

Woo A, Tong T, Culler D (2003) Taming the underlying challenges of reliable multihop routing in sensornetworks. In: First ACM conference on embedded networked sensor systems (SenSys 2003), Novem-ber 2003

Xing G, Lu C, Pless R, Huang Q, (2004) On greedy geographic routing algorithms in sensing-coverednetworks. In: MobiHoc’04, 2004

Yan T, He T, Stankovic J (2003) Differentiated surveillance service for sensor networks. In: First ACMconference on embedded networked sensor systems (SenSys 2003), November 2003

Ye F, Luo H, Cheng J, Lu S, Zhang L (2002a) A two-tier data dissemination model for large-scale wirelesssensor networks. In: 7th ACM MOBICOM conference, 2002

Ye W, Heidemann J, Estrin D (2002b) An energy-efficient mac protocol for wireless sensor networks. In:INFOCOM, 2002

Ye F, Zhong G, Lu S, Zhang L (2003) PEAS: a robust energy conserving protocol for long-lived sensornetworks. In: Proc. of international conference on distributed computing systems (ICDCS), May 2003

Zeng X, Bagrodia R, Gerla M (1998) Glomosim: a library for parallel simulation of large-scale wirelessnetworks. In: The 12th workshop on parallel and distributed simulations, 1998

Zhao J, Govindan R (2003) Understanding packet delivery performance in dense wireless sensor networks.In: Sensys 03, 2003

Zhou G, He T, Stankovic JA (2004) Impact of radio irregularity on wireless sensor networks. In: Thesecond international conference on mobile systems, applications, and services (MobiSys), June 2004

Zorzi M, Rao RR (2003) Geographic random fowarding (GeRaF) for ad hoc and sensor networks: multihopperformance. IEEE Trans Mobile Comput 2(4)

288 Real-Time Syst (2007) 37: 261–289

Tian He is currently an assistant professor in Department of Computer Scienceand Engineering at University of Minnesota-Twin City. He received the Ph.D. de-gree under Professor John A. Stankovic from the University of Virginia, Virginiain 2004, and the M.S. degree from the Institute of Computing Technology, Chi-nese Academy of Sciences, China, in 2000, and the B.S. degree from the NanjingUniversity of Science & Technology, Nanjing, China in 1996. Dr. He is author andco-author of more than 50 refereed publications at international conferences andjournals, including the best paper awards at the second International Conferenceon Mobile Ad-hoc and Sensor Networks and the Fourth ACM Workshop on Secu-rity of Ad Hoc and Sensor Networks. His research interests include wireless sensornetworks, real-time embedded systems and distributed systems. He is a member

of ACM and IEEE.

Brian M. Blum is a Ph.D. student in Department of Computer Science at Univer-sity of Virginia. He is author and co-author of more than 10 refereed publicationsat international conferences and journals. His research interests include wirelesssensor networks, real-time and embedded systems, and distributed Programming,artificial Intelligence, and networking.

Qing Cao is currently a Ph.D. student in Department of Computer Science atUniversity of Illinois, Urbana-Champaign. He received the M.S. degree from theUniversity of Virginia, Virginia, U.S. in 2004, and the B.S. degree from FudanUniversity, Shanghai, China in 2002, both in computer science. His research in-terests include wireless sensor networks, real-time and embedded systems, andnetworking systems.

John A. Stankovic received his Ph.D. from Brown University in 1979 and thenserved on the faculty of University of Massachusetts at Amherst. He came to UVaas BP America Professor and Chair of the Department of Computer Science in1997. Professor Stankovic is a Fellow of the IEEE, a Fellow of the ACM, and heserved on the Computing Research Association Board of Directors for nine years.He received an IEEE Award for Outstanding Technical Contributions and Leader-ship in Real-Time Systems, and an Outstanding Scholar Award from the Univer-sity of Massachusetts. He was Co-Founder and Co-Editor-in-Chief of the Inter-national Journal on Real-Time Systems, Editor-in-Chief of IEEE Transactions onParallel and Distributed Computing, Associate Editor for ACM Transactions onWireless Sensor Networks, Associate Editor for ACM Transactions on Embedded

Systems, and Book Series Editor for Real-Time Systems.

Real-Time Syst (2007) 37: 261–289 289

Sang H. Son is a Professor at the Department of Computer Science of Universityof Virginia. He received the B.Sc. degree in electronics engineering from SeoulNational University, M.Sc. degree from KAIST, and the Ph.D. in computer sciencefrom University of Maryland, College Park in 1986. He has been a Visiting Pro-fessor at KAIST, City University of HongKong, Ecole Centrale de Lille in France,Linköping University, and University of Skövde in Sweden. His current researchinterests include real-time computing, data services, QoS management, wirelesssensor networks, and information security. He has served as an Associate Editorof IEEE Transactions on Parallel and Distributed Systems, and is currently serv-ing as an Associate Editor for Real-Time Systems Journal and Journal of BusinessPerformance Management. He has been on the executive board of the IEEE TC

on Real-Time Systems, and served as the Program Chair or General Chair of several real-time and data-base conferences, including IEEE Real-Time Systems Symposium and IEEE Conference on ElectronicCommerce. He received the Outstanding Contribution Award at the IEEE Conference on Embedded andReal-Time Computing Systems and Applications in 2004.

Tarek F. Abdelzaher his Ph.D. from the University of Michigan in 1999 on Qual-ity of Service Adaptation in Real-Time Systems. He has been an Assistant Pro-fessor at the University of Virginia, where he founded the Software PredictabilityGroup, until his promotion with tenure in 2005. He is currently an Associate Pro-fessor at the Department of Computer Science, the University of Illinois at UrbanaChampaign. He has authored/coauthored three book chapters and more than 60refereed publications in leading conferences and journals in several fields includ-ing real-time computing, distributed systems, sensor networks, and control. He isEditor-in-Chief of the Journal of Real-Time Systems, an Associate Editor of theIEEE Transactions on Mobile Computing, the ACM Transaction on Sensor Net-works, the International Journal of Embedded Systems and the Ad Hoc Networks

Journal, as well as Editor of ACM SIGBED Review. He was Guest Editor for the Journal of ComputerCommunications and the Journal of Real-Time Systems, and is Co-Editor of IEEE Distributed SystemsOnline. He served on numerous technical program committees in real-time computing, networking, qual-ity of service, distributed systems, sensor networks, multimedia, and mobile computing, among others. Healso held several conference organization positions including Program Chair of RTAS 2004, Finance Chairof IPSN 2006, Sensor Networks Vice Chair of RTSS 2005 and ICDCS 2006, System Vice-Chair of DCoSS2006, and General Chair of RTAS 2005. He is currently Finance Chair of Sensys 2006, Program Chair ofRTSS 2006 and General Chair of IPSN 2007. Abdelzaher’s research interests lie broadly in understandingand controlling the temporal properties of software systems in the face of increasing complexity, distri-bution, and degree of embedding in an external physical environment. Tarek Abdelzaher is a member ofIEEE and ACM.